ROUTEBusiness workflows
My X feed is full of posts on Jev by @typesafeai and as a Growth Marketer i've been really confu
Arya Chandra@AryacChandra
SignalChoice - picks from a set of options.
Pattern↑ Jev AI decision models
My X feed is full of posts on Jev by @typesafeai and as a Growth Marketer i've been really confused about what it really is.
So I decided to study up on it - here are my notes:
> Jev is not an LLM. It is a decision classifier. For eg: if you point it to a Basketball and give it 3 options:
1. Orange
2. Yellow
3. Green
and ask it "Hey Jev what is the colour of this basketball?" It will return a confidence score for each colour like so:
Orange: 90%
Yellow: 5%
Green 5%
now, the options that you just sent it is called 'schema'.
> Jev can answer 3 question types:
1. Choice - picks from a set of options. Returns a probability for each option and an overall confidence score.
2. Score - rates an input against ordered levels, such as low, medium, and high. Returns a continuous score, the underlying distribution, and a confidence value.
3. Noul - answers a yes or no question. Returns the probability that a statement is true.
> traditional autoregressive models generate tokens sequentially but Jev can sample parallely. Now what does that mean:
Lets say you ask ChatGPT 4 questions regarding a support ticket :
1. Is it urgent? 2. Should we escalate? 3. Which department? 4. Is a refund needed?
ChatGPT will respond to you one by one:
Is it Urgent -> Yes -> Should we escalate -> Maybe -> Which department -> Billing -> Is refund need? -> Maybe
Jev will answer all 4 questions parallely because of the scoring system:
Customer message
│
├── Urgent? → 94%
├── Escalate? → 81%
├── Department? → Billing (97%)
└── Refund needed? → 89%
> What are the Use Cases of Jev: Typesafe says it is most appropriate for cases where decisions are required instead of open ended generation.
Imagine the support ticket use case:
It can classify a customer message into these categories and provide further direction for a traditional LLM like GPT to take over. Like if Jev decides an immediate reply is required (99.8%) it gets passed to GPT to generate the reply in text form to be sent to the customer.
Another use case I can think of is lead classification: You get many inbound leads through your form -> Jev decides which leads are high intent leads and passes it to GPT -> GPT can then send an email to the high intent lead.